| 1 |
//===-- sanitizer_procmaps_mac.cpp ----------------------------------------===// |
| 2 |
// |
| 3 |
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
// See https://llvm.org/LICENSE.txt for license information. |
| 5 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
// |
| 7 |
//===----------------------------------------------------------------------===// |
| 8 |
// |
| 9 |
// Information about the process mappings (Mac-specific parts). |
| 10 |
//===----------------------------------------------------------------------===// |
| 11 |
|
| 12 |
#include "sanitizer_platform.h" |
| 13 |
#if SANITIZER_APPLE |
| 14 |
#include "sanitizer_common.h" |
| 15 |
#include "sanitizer_placement_new.h" |
| 16 |
#include "sanitizer_procmaps.h" |
| 17 |
|
| 18 |
#include <mach-o/dyld.h> |
| 19 |
#include <mach-o/loader.h> |
| 20 |
#include <mach/mach.h> |
| 21 |
|
| 22 |
// These are not available in older macOS SDKs. |
| 23 |
#ifndef CPU_SUBTYPE_X86_64_H |
| 24 |
#define CPU_SUBTYPE_X86_64_H ((cpu_subtype_t)8) /* Haswell */ |
| 25 |
#endif |
| 26 |
#ifndef CPU_SUBTYPE_ARM_V7S |
| 27 |
#define CPU_SUBTYPE_ARM_V7S ((cpu_subtype_t)11) /* Swift */ |
| 28 |
#endif |
| 29 |
#ifndef CPU_SUBTYPE_ARM_V7K |
| 30 |
#define CPU_SUBTYPE_ARM_V7K ((cpu_subtype_t)12) |
| 31 |
#endif |
| 32 |
#ifndef CPU_TYPE_ARM64 |
| 33 |
#define CPU_TYPE_ARM64 (CPU_TYPE_ARM | CPU_ARCH_ABI64) |
| 34 |
#endif |
| 35 |
|
| 36 |
namespace __sanitizer { |
| 37 |
|
| 38 |
// Contains information used to iterate through sections. |
| 39 |
struct MemoryMappedSegmentData { |
| 40 |
char name[kMaxSegName]; |
| 41 |
uptr nsects; |
| 42 |
const char *current_load_cmd_addr; |
| 43 |
u32 lc_type; |
| 44 |
uptr base_virt_addr; |
| 45 |
uptr addr_mask; |
| 46 |
}; |
| 47 |
|
| 48 |
template <typename Section> |
| 49 |
static void NextSectionLoad(LoadedModule *module, MemoryMappedSegmentData *data, |
| 50 |
bool isWritable) { |
| 51 |
const Section *sc = (const Section *)data->current_load_cmd_addr; |
| 52 |
data->current_load_cmd_addr += sizeof(Section); |
| 53 |
|
| 54 |
uptr sec_start = (sc->addr & data->addr_mask) + data->base_virt_addr; |
| 55 |
uptr sec_end = sec_start + sc->size; |
| 56 |
module->addAddressRange(sec_start, sec_end, /*executable=*/false, isWritable, |
| 57 |
sc->sectname); |
| 58 |
} |
| 59 |
|
| 60 |
void MemoryMappedSegment::AddAddressRanges(LoadedModule *module) { |
| 61 |
// Don't iterate over sections when the caller hasn't set up the |
| 62 |
// data pointer, when there are no sections, or when the segment |
| 63 |
// is executable. Avoid iterating over executable sections because |
| 64 |
// it will confuse libignore, and because the extra granularity |
| 65 |
// of information is not needed by any sanitizers. |
| 66 |
if (!data_ || !data_->nsects || IsExecutable()) { |
| 67 |
module->addAddressRange(start, end, IsExecutable(), IsWritable(), |
| 68 |
data_ ? data_->name : nullptr); |
| 69 |
return; |
| 70 |
} |
| 71 |
|
| 72 |
do { |
| 73 |
if (data_->lc_type == LC_SEGMENT) { |
| 74 |
NextSectionLoad<struct section>(module, data_, IsWritable()); |
| 75 |
#ifdef MH_MAGIC_64 |
| 76 |
} else if (data_->lc_type == LC_SEGMENT_64) { |
| 77 |
NextSectionLoad<struct section_64>(module, data_, IsWritable()); |
| 78 |
#endif |
| 79 |
} |
| 80 |
} while (--data_->nsects); |
| 81 |
} |
| 82 |
|
| 83 |
MemoryMappingLayout::MemoryMappingLayout(bool cache_enabled) { |
| 84 |
Reset(); |
| 85 |
} |
| 86 |
|
| 87 |
MemoryMappingLayout::~MemoryMappingLayout() { |
| 88 |
} |
| 89 |
|
| 90 |
bool MemoryMappingLayout::Error() const { |
| 91 |
return false; |
| 92 |
} |
| 93 |
|
| 94 |
// More information about Mach-O headers can be found in mach-o/loader.h |
| 95 |
// Each Mach-O image has a header (mach_header or mach_header_64) starting with |
| 96 |
// a magic number, and a list of linker load commands directly following the |
| 97 |
// header. |
| 98 |
// A load command is at least two 32-bit words: the command type and the |
| 99 |
// command size in bytes. We're interested only in segment load commands |
| 100 |
// (LC_SEGMENT and LC_SEGMENT_64), which tell that a part of the file is mapped |
| 101 |
// into the task's address space. |
| 102 |
// The |vmaddr|, |vmsize| and |fileoff| fields of segment_command or |
| 103 |
// segment_command_64 correspond to the memory address, memory size and the |
| 104 |
// file offset of the current memory segment. |
| 105 |
// Because these fields are taken from the images as is, one needs to add |
| 106 |
// _dyld_get_image_vmaddr_slide() to get the actual addresses at runtime. |
| 107 |
|
| 108 |
void MemoryMappingLayout::Reset() { |
| 109 |
// Count down from the top. |
| 110 |
// TODO(glider): as per man 3 dyld, iterating over the headers with |
| 111 |
// _dyld_image_count is thread-unsafe. We need to register callbacks for |
| 112 |
// adding and removing images which will invalidate the MemoryMappingLayout |
| 113 |
// state. |
| 114 |
data_.current_image = _dyld_image_count(); |
| 115 |
data_.current_load_cmd_count = -1; |
| 116 |
data_.current_load_cmd_addr = 0; |
| 117 |
data_.current_magic = 0; |
| 118 |
data_.current_filetype = 0; |
| 119 |
data_.current_arch = kModuleArchUnknown; |
| 120 |
internal_memset(data_.current_uuid, 0, kModuleUUIDSize); |
| 121 |
} |
| 122 |
|
| 123 |
// The dyld load address should be unchanged throughout process execution, |
| 124 |
// and it is expensive to compute once many libraries have been loaded, |
| 125 |
// so cache it here and do not reset. |
| 126 |
static mach_header *dyld_hdr = 0; |
| 127 |
static const char kDyldPath[] = "/usr/lib/dyld"; |
| 128 |
static const int kDyldImageIdx = -1; |
| 129 |
|
| 130 |
// static |
| 131 |
void MemoryMappingLayout::CacheMemoryMappings() { |
| 132 |
// No-op on Mac for now. |
| 133 |
} |
| 134 |
|
| 135 |
void MemoryMappingLayout::LoadFromCache() { |
| 136 |
// No-op on Mac for now. |
| 137 |
} |
| 138 |
|
| 139 |
static bool IsDyldHdr(const mach_header *hdr) { |
| 140 |
return (hdr->magic == MH_MAGIC || hdr->magic == MH_MAGIC_64) && |
| 141 |
hdr->filetype == MH_DYLINKER; |
| 142 |
} |
| 143 |
|
| 144 |
// _dyld_get_image_header() and related APIs don't report dyld itself. |
| 145 |
// We work around this by manually recursing through the memory map |
| 146 |
// until we hit a Mach header matching dyld instead. These recurse |
| 147 |
// calls are expensive, but the first memory map generation occurs |
| 148 |
// early in the process, when dyld is one of the only images loaded, |
| 149 |
// so it will be hit after only a few iterations. These assumptions don't |
| 150 |
// hold on macOS 13+ anymore (dyld itself has moved into the shared cache). |
| 151 |
|
| 152 |
// FIXME: Unfortunately, the upstream revised version to deal with macOS 13+ |
| 153 |
// is incompatible with GCC and also uses APIs not available on earlier |
| 154 |
// systems which we support; backed out for now. |
| 155 |
|
| 156 |
static mach_header *GetDyldImageHeaderViaVMRegion() { |
| 157 |
vm_address_t address = 0; |
| 158 |
|
| 159 |
while (true) { |
| 160 |
vm_size_t size = 0; |
| 161 |
unsigned depth = 1; |
| 162 |
struct vm_region_submap_info_64 info; |
| 163 |
mach_msg_type_number_t count = VM_REGION_SUBMAP_INFO_COUNT_64; |
| 164 |
kern_return_t err = |
| 165 |
vm_region_recurse_64(mach_task_self(), &address, &size, &depth, |
| 166 |
(vm_region_info_t)&info, &count); |
| 167 |
if (err != KERN_SUCCESS) return nullptr; |
| 168 |
|
| 169 |
if (size >= sizeof(mach_header) && info.protection & kProtectionRead) { |
| 170 |
mach_header *hdr = (mach_header *)address; |
| 171 |
if (IsDyldHdr(hdr)) { |
| 172 |
return hdr; |
| 173 |
} |
| 174 |
} |
| 175 |
address += size; |
| 176 |
} |
| 177 |
} |
| 178 |
|
| 179 |
const mach_header *get_dyld_hdr() { |
| 180 |
if (!dyld_hdr) { |
| 181 |
// On macOS 13+, dyld itself has moved into the shared cache. Looking it up |
| 182 |
// via vm_region_recurse_64() causes spins/hangs/crashes. |
| 183 |
// FIXME: find a way to do this compatible with GCC. |
| 184 |
if (GetMacosAlignedVersion() >= MacosVersion(13, 0)) { |
| 185 |
VReport(1, |
| 186 |
"looking up the dyld image header in the shared cache on " |
| 187 |
"macOS 13+ is not yet supported. Falling back to " |
| 188 |
"lookup via vm_region_recurse_64().\n"); |
| 189 |
dyld_hdr = GetDyldImageHeaderViaVMRegion(); |
| 190 |
} else { |
| 191 |
dyld_hdr = GetDyldImageHeaderViaVMRegion(); |
| 192 |
} |
| 193 |
CHECK(dyld_hdr); |
| 194 |
} |
| 195 |
|
| 196 |
return dyld_hdr; |
| 197 |
} |
| 198 |
|
| 199 |
// Next and NextSegmentLoad were inspired by base/sysinfo.cc in |
| 200 |
// Google Perftools, https://github.com/gperftools/gperftools. |
| 201 |
|
| 202 |
// NextSegmentLoad scans the current image for the next segment load command |
| 203 |
// and returns the start and end addresses and file offset of the corresponding |
| 204 |
// segment. |
| 205 |
// Note that the segment addresses are not necessarily sorted. |
| 206 |
template <u32 kLCSegment, typename SegmentCommand> |
| 207 |
static bool NextSegmentLoad(MemoryMappedSegment *segment, |
| 208 |
MemoryMappedSegmentData *seg_data, |
| 209 |
MemoryMappingLayoutData *layout_data) { |
| 210 |
const char *lc = layout_data->current_load_cmd_addr; |
| 211 |
|
| 212 |
layout_data->current_load_cmd_addr += ((const load_command *)lc)->cmdsize; |
| 213 |
layout_data->current_load_cmd_count--; |
| 214 |
if (((const load_command *)lc)->cmd == kLCSegment) { |
| 215 |
const SegmentCommand* sc = (const SegmentCommand *)lc; |
| 216 |
uptr base_virt_addr, addr_mask; |
| 217 |
if (layout_data->current_image == kDyldImageIdx) { |
| 218 |
base_virt_addr = (uptr)get_dyld_hdr(); |
| 219 |
// vmaddr is masked with 0xfffff because on macOS versions < 10.12, |
| 220 |
// it contains an absolute address rather than an offset for dyld. |
| 221 |
// To make matters even more complicated, this absolute address |
| 222 |
// isn't actually the absolute segment address, but the offset portion |
| 223 |
// of the address is accurate when combined with the dyld base address, |
| 224 |
// and the mask will give just this offset. |
| 225 |
addr_mask = 0xfffff; |
| 226 |
} else { |
| 227 |
base_virt_addr = |
| 228 |
(uptr)_dyld_get_image_vmaddr_slide(layout_data->current_image); |
| 229 |
addr_mask = ~0; |
| 230 |
} |
| 231 |
|
| 232 |
segment->start = (sc->vmaddr & addr_mask) + base_virt_addr; |
| 233 |
segment->end = segment->start + sc->vmsize; |
| 234 |
// Most callers don't need section information, so only fill this struct |
| 235 |
// when required. |
| 236 |
if (seg_data) { |
| 237 |
seg_data->nsects = sc->nsects; |
| 238 |
seg_data->current_load_cmd_addr = |
| 239 |
(const char *)lc + sizeof(SegmentCommand); |
| 240 |
seg_data->lc_type = kLCSegment; |
| 241 |
seg_data->base_virt_addr = base_virt_addr; |
| 242 |
seg_data->addr_mask = addr_mask; |
| 243 |
internal_strncpy(seg_data->name, sc->segname, |
| 244 |
ARRAY_SIZE(seg_data->name)); |
| 245 |
} |
| 246 |
|
| 247 |
// Return the initial protection. |
| 248 |
segment->protection = sc->initprot; |
| 249 |
segment->offset = (layout_data->current_filetype == |
| 250 |
/*MH_EXECUTE*/ 0x2) |
| 251 |
? sc->vmaddr |
| 252 |
: sc->fileoff; |
| 253 |
if (segment->filename) { |
| 254 |
const char *src = (layout_data->current_image == kDyldImageIdx) |
| 255 |
? kDyldPath |
| 256 |
: _dyld_get_image_name(layout_data->current_image); |
| 257 |
internal_strncpy(segment->filename, src, segment->filename_size); |
| 258 |
} |
| 259 |
segment->arch = layout_data->current_arch; |
| 260 |
internal_memcpy(segment->uuid, layout_data->current_uuid, kModuleUUIDSize); |
| 261 |
return true; |
| 262 |
} |
| 263 |
return false; |
| 264 |
} |
| 265 |
|
| 266 |
ModuleArch ModuleArchFromCpuType(cpu_type_t cputype, cpu_subtype_t cpusubtype) { |
| 267 |
cpusubtype = cpusubtype & ~CPU_SUBTYPE_MASK; |
| 268 |
switch (cputype) { |
| 269 |
case CPU_TYPE_I386: |
| 270 |
return kModuleArchI386; |
| 271 |
case CPU_TYPE_X86_64: |
| 272 |
if (cpusubtype == CPU_SUBTYPE_X86_64_ALL) return kModuleArchX86_64; |
| 273 |
if (cpusubtype == CPU_SUBTYPE_X86_64_H) return kModuleArchX86_64H; |
| 274 |
CHECK(0 && "Invalid subtype of x86_64"); |
| 275 |
return kModuleArchUnknown; |
| 276 |
case CPU_TYPE_ARM: |
| 277 |
if (cpusubtype == CPU_SUBTYPE_ARM_V6) return kModuleArchARMV6; |
| 278 |
if (cpusubtype == CPU_SUBTYPE_ARM_V7) return kModuleArchARMV7; |
| 279 |
if (cpusubtype == CPU_SUBTYPE_ARM_V7S) return kModuleArchARMV7S; |
| 280 |
if (cpusubtype == CPU_SUBTYPE_ARM_V7K) return kModuleArchARMV7K; |
| 281 |
CHECK(0 && "Invalid subtype of ARM"); |
| 282 |
return kModuleArchUnknown; |
| 283 |
case CPU_TYPE_ARM64: |
| 284 |
return kModuleArchARM64; |
| 285 |
default: |
| 286 |
CHECK(0 && "Invalid CPU type"); |
| 287 |
return kModuleArchUnknown; |
| 288 |
} |
| 289 |
} |
| 290 |
|
| 291 |
static const load_command *NextCommand(const load_command *lc) { |
| 292 |
return (const load_command *)((const char *)lc + lc->cmdsize); |
| 293 |
} |
| 294 |
|
| 295 |
static void FindUUID(const load_command *first_lc, u8 *uuid_output) { |
| 296 |
for (const load_command *lc = first_lc; lc->cmd != 0; lc = NextCommand(lc)) { |
| 297 |
if (lc->cmd != LC_UUID) continue; |
| 298 |
|
| 299 |
const uuid_command *uuid_lc = (const uuid_command *)lc; |
| 300 |
const uint8_t *uuid = &uuid_lc->uuid[0]; |
| 301 |
internal_memcpy(uuid_output, uuid, kModuleUUIDSize); |
| 302 |
return; |
| 303 |
} |
| 304 |
} |
| 305 |
|
| 306 |
static bool IsModuleInstrumented(const load_command *first_lc) { |
| 307 |
for (const load_command *lc = first_lc; lc->cmd != 0; lc = NextCommand(lc)) { |
| 308 |
if (lc->cmd != LC_LOAD_DYLIB) continue; |
| 309 |
|
| 310 |
const dylib_command *dylib_lc = (const dylib_command *)lc; |
| 311 |
uint32_t dylib_name_offset = dylib_lc->dylib.name.offset; |
| 312 |
const char *dylib_name = ((const char *)dylib_lc) + dylib_name_offset; |
| 313 |
dylib_name = StripModuleName(dylib_name); |
| 314 |
if (dylib_name != 0 && (internal_strstr(dylib_name, "libclang_rt."))) { |
| 315 |
return true; |
| 316 |
} |
| 317 |
} |
| 318 |
return false; |
| 319 |
} |
| 320 |
|
| 321 |
const ImageHeader *MemoryMappingLayout::CurrentImageHeader() { |
| 322 |
const mach_header *hdr = (data_.current_image == kDyldImageIdx) |
| 323 |
? get_dyld_hdr() |
| 324 |
: _dyld_get_image_header(data_.current_image); |
| 325 |
return (const ImageHeader *)hdr; |
| 326 |
} |
| 327 |
|
| 328 |
bool MemoryMappingLayout::Next(MemoryMappedSegment *segment) { |
| 329 |
for (; data_.current_image >= kDyldImageIdx; data_.current_image--) { |
| 330 |
const mach_header *hdr = (const mach_header *)CurrentImageHeader(); |
| 331 |
if (!hdr) continue; |
| 332 |
if (data_.current_load_cmd_count < 0) { |
| 333 |
// Set up for this image; |
| 334 |
data_.current_load_cmd_count = hdr->ncmds; |
| 335 |
data_.current_magic = hdr->magic; |
| 336 |
data_.current_filetype = hdr->filetype; |
| 337 |
data_.current_arch = ModuleArchFromCpuType(hdr->cputype, hdr->cpusubtype); |
| 338 |
switch (data_.current_magic) { |
| 339 |
#ifdef MH_MAGIC_64 |
| 340 |
case MH_MAGIC_64: { |
| 341 |
data_.current_load_cmd_addr = |
| 342 |
(const char *)hdr + sizeof(mach_header_64); |
| 343 |
break; |
| 344 |
} |
| 345 |
#endif |
| 346 |
case MH_MAGIC: { |
| 347 |
data_.current_load_cmd_addr = (const char *)hdr + sizeof(mach_header); |
| 348 |
break; |
| 349 |
} |
| 350 |
default: { |
| 351 |
continue; |
| 352 |
} |
| 353 |
} |
| 354 |
FindUUID((const load_command *)data_.current_load_cmd_addr, |
| 355 |
data_.current_uuid); |
| 356 |
data_.current_instrumented = IsModuleInstrumented( |
| 357 |
(const load_command *)data_.current_load_cmd_addr); |
| 358 |
} |
| 359 |
|
| 360 |
while (data_.current_load_cmd_count > 0) { |
| 361 |
switch (data_.current_magic) { |
| 362 |
// data_.current_magic may be only one of MH_MAGIC, MH_MAGIC_64. |
| 363 |
#ifdef MH_MAGIC_64 |
| 364 |
case MH_MAGIC_64: { |
| 365 |
if (NextSegmentLoad<LC_SEGMENT_64, struct segment_command_64>( |
| 366 |
segment, segment->data_, &data_)) |
| 367 |
return true; |
| 368 |
break; |
| 369 |
} |
| 370 |
#endif |
| 371 |
case MH_MAGIC: { |
| 372 |
if (NextSegmentLoad<LC_SEGMENT, struct segment_command>( |
| 373 |
segment, segment->data_, &data_)) |
| 374 |
return true; |
| 375 |
break; |
| 376 |
} |
| 377 |
} |
| 378 |
} |
| 379 |
// If we get here, no more load_cmd's in this image talk about |
| 380 |
// segments. Go on to the next image. |
| 381 |
data_.current_load_cmd_count = -1; // This will trigger loading next image |
| 382 |
} |
| 383 |
return false; |
| 384 |
} |
| 385 |
|
| 386 |
void MemoryMappingLayout::DumpListOfModules( |
| 387 |
InternalMmapVectorNoCtor<LoadedModule> *modules) { |
| 388 |
Reset(); |
| 389 |
InternalMmapVector<char> module_name(kMaxPathLength); |
| 390 |
MemoryMappedSegment segment(module_name.data(), module_name.size()); |
| 391 |
MemoryMappedSegmentData data; |
| 392 |
segment.data_ = &data; |
| 393 |
while (Next(&segment)) { |
| 394 |
// skip the __PAGEZERO segment, its vmsize is 0 |
| 395 |
if (segment.filename[0] == '\0' || (segment.start == segment.end)) |
| 396 |
continue; |
| 397 |
LoadedModule *cur_module = nullptr; |
| 398 |
if (!modules->empty() && |
| 399 |
0 == internal_strcmp(segment.filename, modules->back().full_name())) { |
| 400 |
cur_module = &modules->back(); |
| 401 |
} else { |
| 402 |
modules->push_back(LoadedModule()); |
| 403 |
cur_module = &modules->back(); |
| 404 |
cur_module->set(segment.filename, segment.start, segment.arch, |
| 405 |
segment.uuid, data_.current_instrumented); |
| 406 |
} |
| 407 |
segment.AddAddressRanges(cur_module); |
| 408 |
} |
| 409 |
} |
| 410 |
|
| 411 |
} // namespace __sanitizer |
| 412 |
|
| 413 |
#endif // SANITIZER_APPLE |